Kokoro-82M
A small open-weight text-to-speech model for local speech generation with preset voice packs.
Score calculation
(4 + 5 + 8 + 8 + 1 + 7) ÷ 6 = 5.5
6 of six dimensions rated. Unrated dimensions are excluded from the average.
Full scoring method ↗Setup details
Upstream hexgrad Kokoro-82M model card, weights and related voice-pack release.
What fits your needs
Upstream hexgrad Kokoro-82M model card, weights and related voice-pack release.
What it does
A small open-weight text-to-speech model for local speech generation with preset voice packs.
See the evidence
Route
Downloadable Apache-2.0 weights; choose a compatible local runtime and separately check each voice pack.
Tradeoffs to expect
Terms and data
The weights can run locally; review each voice pack and dependency license separately.
See the evidence
How to switch
Run locally
Download the model and runtime from the official source, then evaluate language, voice and hardware fit.
See the evidence
Practical questions
Rights and permissions
Check the dataset and voice-specific licenses for the exact model files and runtime you download.
Data path
With a local runtime, audio need not leave your device; confirm app telemetry, logs, backups and any remote integrations separately.
See the evidence
The model card reports training compute and provides a small local model; electricity and inference energy remain unmeasured.
See the evidence
Kokoro v1.0 reports about 1,000 A100 GPU-hours across its training releases and publishes an 82M-parameter model that can run locally. The grade credits scoped compute reporting and a practical runtime choice. GPU-hours are not kWh, and neither local inference nor a small model proves low emissions.
1 + 2 reporting + 0 electricity + 1 controls = 4/10. This grades documented practices and transparency.
Comparable per-minute or per-generation electricity, water use, lifecycle emissions and independently verified clean-power sourcing are not published in these sources.
- Reporting2/3
- The v1.0 model card reports a dated training release and about 1,000 A100 GPU-hours in total. This is a scoped compute disclosure, not metered electricity or an inference footprint.
- Electricity sourcing0/3
- No attributable clean-electricity procurement or hourly matching evidence was found in the cited primary material.
- Resource controls1/3
- The downloadable 82M-parameter model provides a small, locally controllable inference option. The card does not measure energy savings against another system.
Water disclosure
Water reporting has not been assessed for this choice.
The model card discloses training source classes but does not publish a complete rights ledger.
See the evidence
The upstream Kokoro model card says the model used permissive or non-copyrighted audio, including public-domain, permissively licensed and synthetic speech. It names two Creative Commons source datasets and their approximate durations, but does not itemize the rest of the training data. This is model-specific provenance and a meaningful rights claim, not a complete independent permissions audit.
1 + 2 source transparency + 2 permission + 0 creator controls = 5/10, with the permission cap applied.
The full source inventory, speaker consent records, exact licenses for the remaining training data, rights-holder removal process and independent rights audit are not published in the reviewed model materials.
- Source transparency2/3
- The card names broad source classes and two identified datasets, but not the full training data.
- Documented permission2/4
- The publisher says the training data uses permissive/non-copyrighted audio and identifies named CC-licensed components; the material remainder is not itemized.
- Creator controls0/2
- No rights-holder opt-out, removal or retraining remedy is documented in the model release sources.
Local inference can keep the audio and transcript on your own device.
See the evidence
The published model can be downloaded and run locally. Once downloaded, inference need not send recordings or transcripts to the model publisher; software dependencies, telemetry, backups and any remote application remain separate data paths under the operator’s control.
The model publisher does not audit the user’s runtime, network connections, logging, backups or downstream application. Verify those components for your installation.
The model weights and code are published under a permissive license; training remains difficult to reproduce.
See the evidence
Kokoro-82M weights are published on the upstream Hugging Face repository under Apache-2.0, with code in the linked official repository. This enables local use subject to the included license. A complete training dataset and reproducible end-to-end recipe are not supplied.
A complete data manifest, exact training recipe and independently reproduced end-to-end training are not supplied.
hexgrad publishes public company and careers information, but the reviewed pages don’t show worker protections across its AI supply chain.
See the evidence
The reviewed company and product material provides some public visibility into the organization, but not independently verified evidence on annotation, moderation, contractor pay, worker representation, grievance handling or remedies. This score reflects what the vendor documents, not a finding that workers are mistreated.
1 + 0 + 0 + 0 = 1/10. Documented worker protections; missing disclosure is an evidence gap.
Contractor conditions, worker voice and remedy, supply-chain coverage and independent worker-level assurance are not established in these sources.
- Pay & conditions0/3
- No clear public evidence on pay or working conditions across employees and contractors was found in the reviewed company and product pages.
- Worker voice & remedy0/3
- No worker grievance, representation or remedy process was found in the reviewed public pages.
- Evidence & follow-through0/3
- No independent worker-level implementation audit was found in the reviewed public pages.
No military customer, contract or operational defense offering was established in the official sources reviewed for hexgrad.
See the evidence
The reviewed official company and product materials establish an audio product business but do not identify a military customer, defense contract, operational defense offering or military-use restriction. This is a bounded finding about sources checked, not proof that no relationship exists. No identity, location or supplier assumptions are used.
7 base + 0 safeguard adjustment = 7/10 (0 of 9 safeguard points; every 3 adds 1). No military relationship was established in the cited sources. This ranks above known defense ties without claiming a verified clean record.
Named customers, contract scopes, downstream deployments, customer-screening practices and independent military-use oversight may not be disclosed in the reviewed material.
- Use restrictions0/4
- The permissive model/code license states no military-use restrictions; no separate model-specific military policy was found.
- Enforcement & oversight0/3
- No public military customer review or independent defense-use oversight was found in these sources.
- Relationship disclosure0/2
- The reviewed pages do not name a military customer, contract scope or deployment.
No direct military or defense relationship was established in the reviewed official product, company and terms pages. This is a source-limited finding and does not certify that no relationship exists.
Military rubric ↗hexgrad was not named in the two official BNC technology and corporate target lists checked.
See the evidence
The official BNC corporate boycott and pressure guide and its No Tech campaign page do not name hexgrad among the current targets listed there. BDS uses selective, campaign-based targets; absence from these two lists is neither approval nor a broad assessment of company relationships.
This categorical check covers only the two linked BNC lists as reviewed on the date shown; local campaigns and future list updates may differ.